Summary An approach for determining the moments of the annual runoff distribution from those of climatic variables is presented. The derived moment approach provides an alternative for non-stationary annual runoff frequency analysis as the moments of climate variables are more directly predictable by climate change models. A possible change in non-stationary situations is assumed to appear in the moments and not in the type of distribution. An annual rainfall–runoff model is established first. Then the annual runoff moments are derived from the moments of the input variables of the annual rainfall–runoff model (i.e. annual precipitation, potential evapotranspiration and water storage change) by applying a second-order approximation. The approach is tested in 40 watersheds in the Gan River and Han River basins in China. The Gamma distribution is proved to be the appropriate two-parameter distribution for fitting the annual runoff. Seven annual rainfall–runoff models based on the Budyko hypothesis are used in the derivation of the annual runoff moments. The performance of the derived moment approach is good in general and is mainly influenced by the accuracy of the annual rainfall–runoff model and also by the plausibility of the assumed distribution type. The model based on the Schreiber formula performs best in the two basins. The derived moment approach can be used in ungauged watersheds when neglecting the annual water storage change.